Contention and Queueing in an Experimental Multicomputer: Analytical and Simulation-based Results

Abstract:

We use two methods, analytic modeling and event-driven simulation, to
evaluate the performance of communication in the SHRIMP multicomputer under a
variety
of workloads. This work has two purposes: to learn about the behavior
of the SHRIMP machine, and to explore the tradeoffs between analytic
modeling and simulation as performance prediction techniques.
We find that the system's performance depends mostly on the behavior of the
endpoints; most of the message latency is spent arbitrating for the
sender's and receiver's memory buses, and doing DMA transfers of data
to and from memory. The raw speed of the network is relatively
unimportant. Performance is mainly determined by details of DMA
speed, network connectivity, and policies for arbitrating for buses
and network links.
Even for very simple synthetic benchmark programs, the analytic model
and the simulator differ by roughly 10% in their performance
predictions. Although we believe the simulator to be more accurate
than the model, each method has difficulty accurately representing
some feature of the real system. Finally, we observe that each of the
two tools was valuable in helping to improve the accuracy of the other.